Enriching contextualized semantic representation with textual information transmission for COVID-19 fake news detection: A study on English and Persian.

This article focuses on detecting COVID-19 fake news in English and Persian by analyzing textual properties derived from language. It employs a model combining XLM-RoBERTa, a cross-lingual BERT-based representation, with a convolutional neural network (CNN) classifier, enriched by features including information theory metrics (surprisal), readability formulas, stylometry, and linguistic annotations at phonological, morphological, and syntactic levels. Experimental results show that augmenting the basic model with readability and stylometry features yields the best improvement for English fake news detection, while for Persian, combining readability, surprisal, morphological, and stylometry features achieves the highest performance. The study highlights that optimal feature sets differ by language and that morphological information is particularly valuable when annotated data is limited.